Reliability Inference Drives Cue Extraction in Large Language Models Consuming External Reasoning Traces

Large language models increasingly consume externally produced reasoning traces (e.g., tool logs and multi-agent outputs), requiring them to infer trace reliability and arbitrate between independent computation and cue extraction. We introduce a controlled corruption paradigm that independently manipulates trace degradation and embedded answer cues. Across 28 experiments (71,168 trials) with GPT-4o and Claude, we find a model-dependent strategy shift: in verifiable arithmetic, GPT-4o's wrong-cue following increases monotonically with corruption (0.5% to 8.5% at λ=0.8), whereas Claude shows little change. This vulnerability is determined by the consumer model rather than the trace producer, and is amplified by cue salience (up to 21%) and by naturalistic tool-log trace formats (29.1% at λ=0.8—approximately 3× higher than standard formats). Key findings:- Reliability-gated cue extraction: Trace corruption induces graded shift toward answer-cue following (GPT-4o), while Claude remains resistant- Consumer-determined vulnerability: The model reading the trace determines susceptibility, not the producer- Naturalistic format amplification: Tool-log traces produce ~3× higher vulnerability than standard formats- Cue style independence: Implicit cues in natural conclusions are equally effective as explicit markers (p=0.75)- Dual reliability mechanisms: Continuous parameters trigger scrutiny; discrete "Verified" labels trigger trust delegation- Defense effectiveness: Simple verification prompts suppress cue following to 0% Code and data: https://github.com/HIDEKI-SQ/cot-reliability-gating (v4.0.0, DOI: 10.5281/zenodo.18207632)